Open LimiX Studio

LIMIX-2 / CURRENT FLAGSHIP

One model formany structured‑dataquestions

A single pretrained LimiX model predicts outcomes, recovers missing values, and captures relationships across structured data.

Get LimiX-2 View on GitHub Read the report

Current flagship

Meet LimiX-2

One 400M-parameter checkpoint for classification, regression, and missing-value imputation—without task-specific parameter updates.

LIMIX-2 / CONTEXTUAL MECHANISM NETWORK400M PARAMETERS
CONTEXT → QUERYCELL / COLUMN / BLOCK
One pretrained model for classification, regression, and missing-value imputation.
Parameters
406.2M
Checkpoint
One
Released tasks
Three
Released
16 Sep 2026
Inference
CUDA recommended · CPU supported

Released scopeCausal-skeleton recovery is reported as an attention probe. It is not a released predictor task.

Released capabilities

One predictor for many tasks

Released capability

Predict class probabilities for new rows

Condition on labeled context rows to return class probabilities for each query row.

  • Risk
  • Condition
  • Failure
  • Churn
CONTEXT + QUERY
OUTPUT
Released task · one LimiXPredictor interface
01Classification

Released capability

Predict class probabilities for new rows

Condition on labeled context rows to return class probabilities for each query row.

  • Risk
  • Condition
  • Failure
  • Churn
CONTEXT + QUERY
OUTPUT
Released task · one LimiXPredictor interface
02Regression

Released capability

Estimate continuous outcomes from context

Infer a continuous target on its original scale from the surrounding table.

  • Demand
  • Quality
  • Price
  • Performance
CONTEXT + QUERY
OUTPUT
Released task · one LimiXPredictor interface
03Imputation

Released capability

Recover missing values from the surrounding table

Resolve masked cells using the variables and observations that remain visible.

  • Sensor gaps
  • Incomplete records
  • Sparse measurements
CONTEXT + QUERY
OUTPUT
Released task · one LimiXPredictor interface

Method innovation

From a single target to the structure of the table

Model the joint structure

CMNs organize in-context learning around dependencies across variables. Prediction becomes one conditional question among many.
PFN-STYLE OBJECTIVE

Target-centric prediction

p(y | x, Dcontext)
CMN MODELING VIEW

Mechanism-oriented joint modeling

p(x, y | Dcontext)
Simplified from the Technical Report.
01CMNContextual Mechanism Networks

Model the joint structure

CMNs organize in-context learning around dependencies across variables. Prediction becomes one conditional question among many.
PFN-STYLE OBJECTIVE

Target-centric prediction

p(y | x, Dcontext)
CMN MODELING VIEW

Mechanism-oriented joint modeling

p(x, y | Dcontext)
Simplified from the Technical Report.
02CCMMContext-Conditional Masked Modeling

Learn across varied observation patterns

CCMM combines target prediction with masked-feature reconstruction across entry, column, and block masks.
CONDITIONAL TASKRecover isolated hidden entries.
Simplified from the Technical Report.
03SCMSCM-generated pretraining data

Pretrain across diverse synthetic mechanisms

Synthetic pretraining spans diverse structures and observation processes.
01

Causal graph

Diverse DAG structures

02

Functional mechanisms

Linear and nonlinear relations

03

Sample target & features

From the generated SCM

04

Conditional tasks

Prediction and reconstruction

Simplified from the Technical Report.

Reported performance

Measured across broad, real-world benchmarks

TabArena reflects the official live leaderboard accessed 18 Sep 2026. TALENT and BCCO retain the Technical Report values.

TabArena1943Overall Elo

51 datasets · 30 binary · 8 multiclass · 13 regression

TALENT1506Overall Elo

300 total · 12 excluded · 288 evaluated (120 binary · 68 multiclass · 100 regression)

BCCO1432Overall Elo

156 datasets · 71 binary · 35 multiclass · 50 regression

Explore the full benchmark comparison

Designed for real tables

Built for the data you already have

Illustrative table · context and query shown together
RowAgeNUMRegionCATSignalNUMStateCATClassCATScoreNUM
0142North0.812NominalRetain0.08
0267EastReviewAlert0.71
0335West0.925NominalRetain0.03
0451North0.641Alert0.44
Query58South0.603ReviewAlert0.82
Missing Predicted

Keep continuous values and categories together in their native table structure.

Model evolution

The Evolution of LimiX

  1. 01

    Generalize

    LimiX-16M

    16Mparameters

    Established the first generalist LimiX release.

  2. 02

    Compact

    LimiX-2M

    2Mparameters

    Studied compact structured-data modeling and attention bottlenecks.

  3. 03

    Scale

    LimiX-2

    400Mparameters

    Scaled Contextual Mechanism Networks to one checkpoint for three released tasks.

    Latest model

License

Available for research and commercial licensing

MODEL WEIGHTS

Stable AI LimiX Non-Commercial License

Applies to the LimiX-2 model weights. Commercial use requires a separate agreement.

Read weight license
REPOSITORY CODE

Stable AI Technology Co., Ltd. License, Version 1.0

Based on Apache 2.0 with additional attribution and model-naming provisions.

Read code license
COMMERCIAL USE

Separate agreement

Contact Stable AI to discuss commercial licensing and deployment.

Discuss licensing

Start building

From checkpoint to your first prediction

Developer Guide
GitHub Hugging Face Technical Report